Back to skill

Security audit

Save/Load Context

Security checks for vulnerabilities and agentic risk

Overview

This save/load plugin does what it says, but it can expose or replace conversation history because session scoping and saved transcript handling are too loose.

Install only if you are comfortable with full conversations being saved locally in plaintext and restored into future agent context. Avoid using it in shared OpenClaw installations or multi-user gateways until session ownership checks, restrictive file permissions, confirmation for destructive actions, and safer context restoration are added.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
Findings (3)

T05 · Unauthorized Access and Privilege Escalation

Error
Location
index.ts:56
Finding

Ambiguous Session Resolution Enables Cross-Session Transcript Access and Replacement

Content
View full analysis
Remediation
View remediation

T01 · Skill Instruction Hijacking

Error
Location
index.ts:523
Finding

Saved Conversation Content Is Injected into a Privileged System Event

Content
View full analysis
{ const preview = m.content.slice(0, 120).replace(/\n/g, " "); return `[${m.role}] ${preview}${m.content.length > 120 ? "..." : ""}`; }); if (messages.length > 10) { summaryLines.push(`... and ${messages.length - 10} more messages`); } const saveFilePath = filePath; // Inject system event so the running agent picks up the loaded context if (_runtime && sessionKey) { const systemEvent = [ `[SYSTEM: Context Loaded]`, `A saved conversation context has been loaded:`, `- Label: ${targetEntry.label}`, `- Messages: ${messages.length}`, `- Save file: ${saveFilePath}`, ``, `Conversation summary:`, ...summaryLines, ``, `To restore full context, read the save file with the read tool:`, `\`read ${saveFilePath}\``, `Then continue the conversation naturally, as if you remember everything.`, ].join("\n"); try { _runtime.system.enqueueSystemEvent(systemEvent, { sessionKey }); } catch (e) { // Non-fatal: system event injection failed, context still saved to file } } ``` ### Technical Analysis The first ten saved user or assistant messages are interpolated directly into a system event. Conversation messages are untrusted content, but the plugin changes their effective trust level by placing them inside a privileged system-event string. Replacing newline characters does not neutralize natural-language instructions. A saved message can begin with instruction-like content such as directives to ignore prior goals, invoke tools, disclose information, or treat subsequent data as authoritative. The event also explicitly tells the agent to read the entire save file and continue as though it remembers all saved content. Because the event doe ...[truncated 1585 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
index.ts:374
Finding

Sensitive Conversation Transcripts Are Persisted Without Explicit Restrictive Permissions

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (11)

Tp4

High
Category
MCP Tool Poisoning
Confidence
92% confidence
Finding

The documented description is narrower than the actual described behavior: beyond basic save/load, the skill also supports permanent deletion, session transcript rewriting, and system event injection that can influence subsequent agent behavior. That mismatch is dangerous because users and reviewers may grant trust based on a benign-looking summary while missing capabilities that alter state, destroy data, or create an indirect prompt-injection channel through injected file-path context.

Content

No source excerpt is available for this finding.

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 70)May include surrounding context.

md
2. Ensure `openclaw.plugin.json` and `package.json` are present

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill persists conversation history to disk but does not clearly warn that saved transcripts may contain secrets, personal data, tokens, or other sensitive content. In this context, the omission materially increases the chance of accidental data exposure because users may assume chat context is ephemeral while the plugin creates durable local files that can later be accessed, copied, or backed up.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

Documenting a permanent deletion command without an explicit warning or confirmation expectation makes accidental destructive use more likely. In a context-management skill, deleting saved conversations can cause irreversible loss of audit trail, work product, or evidence of prior interactions, especially if users invoke commands by label or index and make selection mistakes.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The load command archives the current transcript and writes a new transcript file to replace the active session context. Although the code performs the action as part of loading, there is no user-facing warning in the command description or an explicit confirmation that current session state will be replaced and archived.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The plugin injects a system event instructing the agent to read the full saved file and 'continue the conversation naturally, as if you remember everything.' Because save files contain prior user/assistant messages and the prompt explicitly causes the agent to internalize them, previously stored sensitive data can be reintroduced into the model context and influence future outputs, potentially disclosing prior conversation content across turns or to the wrong audience if session resolution is incorrect.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The delete path unlinks the save file from disk and removes its index entry, which is a destructive operation. The command metadata and surrounding user guidance do not clearly warn that /load --delete permanently deletes saved context, and there is no confirmation prompt before deletion.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The manifest describes the skill only as "Save and load conversation context to file," but provides no specific trigger phrases, scope limits, or exclusion conditions. In a manifest file, this kind of generic capability description can overlap with common requests about saving or loading context and may cause unintended invocation.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The manifest explicitly says the skill saves and loads conversation context to file, but provides no user-facing warning about persistence, retention, or sensitivity of stored data. Because conversation context can contain secrets, personal data, or proprietary information, silent or poorly disclosed file storage increases the risk of unintended data exposure on disk or later recovery by other users/processes.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
90% confidence
Finding

The skill emits some user-facing messages in Chinese while command descriptions and other responses are in English. This imposes a language choice on users without opt-in or documentation that the skill is intended for a Chinese-speaking context.

Content

No source excerpt is available for this finding.

Unverifiable Dependency: openclaw has 16 known advisory(ies) (CVE-2026-53846 (OpenClaw: Workspace .env npm_execpath could influence bundled runtime dependency); CVE-2026-32064 (OpenClaw's andbox browser noVNC observer lacked VNC authentication); CVE-2026-32006 (OpenClaw has a BlueBubbles group allowlist mismatch via DM pairing-store fallbac) +13 more), but the manifest does not pin a version, so it is unknown whether the installed release is affected

Low
Category
Supply Chain
Confidence
40% confidence
Finding

Dependency has known vulnerabilities (CVEs). Using packages with unpatched security flaws exposes the environment to known exploits.

Content

No source excerpt is available for this finding.